AI for Bio · Model

Enchant v1/v2

Closed multimodal transformer predicting molecular properties and clinically relevant endpoints from early data

Small moleculesAI-native biotechUnited StatesProduct only, no public artifact中文

Overview

NameEnchant v1/v2 · aliases: Enchant v1, Enchant v2, Enchant v3
Full nameThe product name has no expanded form
Ambiguous names (not auto-linked)Enchant
FamilyCovers three generations of one closed product (v1/v2/v3), none with weights or third-party review
  • Enchant v1 ·
  • Enchant v2 ·
  • Enchant v3 ·
OrganizationIambic Therapeutics
Organization typeAI-native biotech
CountryUnited States
Model classProperty prediction / multimodal foundation model
DomainSmall molecules (also: Text, LLMs & agents, Structure & co-folding)
First release · product update
Basis: Company press release for v1 (earliest sourced public evidence)
Note: Product-only: no paper and no weights source
Other milestones
  • First public evidence: · v1 press release source
  • Official announcement: · Company press release and official blog source
Latest update · new version
Basis: v3 launch (official page)
v3 launched: the company describes a 41B-parameter mixture-of-experts multimodal transformer (v1 1B, v2 7B) taking structures, sequences and omics, tabular assays, physics data, text and images, adding inverse design and peptide capabilities; the parameter framing matches the offering document. The developer's 2026-10-16 listing is a corporate event, not a model update
Note: The 41B and mixture-of-experts details were not independently re-verified source
The most recent sourced event, not a claim that nothing changed since.
StatusProduct only, no public artifact · Closed, with no public weights or code; product-only
OpennessCode: Closed
Weights: Closed
All three generations are closed, disclosed only through company pages and press releases, with no weights and no cited API documentation
Benchmarks reported0 counted (unconfirmed 3)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[0], benchmarks_reported[1], benchmarks_reported[2], latest_update.description, training_data.b_private_licensed, training_data.c_own_wetlab_for_training
As of2026-11-16

What it does

Puts molecular structures and several experimental and clinically relevant modalities into one transformer, aiming to close the preclinical-to-clinical data gap and predict ADME, pharmacokinetics and clinically relevant endpoints; v2 claims more than a tenfold scale-up over v1 and v3 adds inverse design, omics and image prediction. The company says it fine-tunes weekly across hundreds of molecular properties. TxGemma has been used as a comparator; TerraBind, backed by the same chip investor, instead trains on public data with its own wet-lab validation.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 0 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
01 CASPUnconfirmed (not counted)—Unconfirmed sourcev3 page claims held-out structures from this round, figure only, no protocol; split from a combined candidate and kept unconfirmed. †disputed · evidence
04 TDC (ADMET + DTI)Unconfirmed (not counted)—Unconfirmed sourceObach human PK result not shown to use the TDC split; three site surfaces unified to unconfirmed †disputed · evidence
15 CAMEOUnconfirmed (not counted)—Unconfirmed sourceSame source; static 2022 set, figure only. Unconfirmed; would count as retrospective use, not participation. †disputed · evidence

Other evaluations (outside the 44)

  • Public human pharmacokinetics dataset (half-life) (Public benchmark not among the 44) Self-reported: The company reports better correlation than the prior best on the full dataset; the announcement links a baseline page without naming a benchmark and does not state the split
  • Low-data setting (Comparison against baselines only) Self-reported: v2 claims that training on 5% of the dataset (about two dozen molecules) beats the prior best trained on all of it
  • Company-internal task sets (Own task set) Self-reported: v2 compares against v1 and TxGemma on a kinase set, one company's ADME set, a cell pharmacodynamics set and rat in-vivo clearance, mostly as figures
  • v3 protein structure quality comparison (Comparison against baselines only) Self-reported: The official page claims better quality than two public protein models on held-out structures, with figures only and no numbers

Developers report: a multimodal transformer pretrained on abundant preclinical data predicts human pharmacokinetics with very little clinical data and beats the prior best in low-data settings.

Independent evidence: No independent evidence found. The model is closed with no weights and every figure is company-reported; no third-party evaluation or replication was found

Training data

Group in Figure 7: Own wet-lab training data.

ClassValueDetail
a. Public databasesYesSaid to come from dozens of external sources, none named; a public human pharmacokinetics dataset is used for the PK task source
b. Private or licensed dataNot disclosedPossibly present but undisclosed; the published site had a claim without scale source
c. Own wet-lab data used for trainingYesSays an in-house automated synthesis and assay platform generates new data continuously, with weekly fine-tuning across dozens of endpoints; scale undisclosed source
d. Wet lab used only for validationNot disclosed

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An AI small-molecule drug discovery company incorporated in the US in 2019 and renamed in 2023-05, headquartered in California with offices in the UK and Ireland; its offering document reports 171 full-time staff as of 2026-09 (140 R&D, 31 administrative). It runs its own pipeline plus a platform.
  • Funding: A $100M Series B in 2023-10 (co-led by two funds, with a chip company and a sequencing company's venture arm as new investors); a $50M extension in 2024-06 (led by sovereign and family capital) taking the round above $150M; a further round of over $100M in 2025-11 with no lead named. The offering document reports about $461.8M raised since inception, $536.5M in total cash and $66.5M of convertible notes converting at listing. The IPO closed 2026-10-16 at $16.00 per share for 9,375,000 shares and $150M.
  • Products and partnerships: The pipeline includes an oral brain-penetrant targeted inhibitor (in a phase 1/1b basket trial as of 2026-09), a programme that filed a clinical trial application on 2026-09-30 and one pre-filing programme; the platform comprises this model and a co-folding model. A 2026-09 collaboration and option-to-license deal with a large pharma company covers two initial targets with an upfront, milestones and tiered royalties, terms undisclosed; a 2026-02 research collaboration and licence with another pharma company covers three initial targets with aggregate milestones above $1.7B and $6.2M of revenue recognised in the first half of 2026; five further pharma and compute partners are named.
  • Related models: TxGemma, TerraBind, NeuralPLexer3

Reading notes

  1. Every figure is company-reported and the model is closed, so nothing can be independently reproduced.
  2. The TDC relation is ruled unconfirmed this round: the blog names no benchmark and no split, and all three published surfaces are unified to unconfirmed.
  3. The human pharmacokinetics correlation first appeared in the v1 announcement; the v2 blog restates the same figure without saying it was re-run, so it is not a v2 result.
  4. The real differentiator is continuously generated in-house wet-lab data in training, but its scale has never been disclosed, and v3's parameters and capabilities rest on company statements only.

Caveats

  1. No counted pairings: all three site relations are unconfirmed.
  2. Closed, with no weights or code and no third-party review.
  3. The scale of in-house wet-lab training data is undisclosed.
  4. Pretraining token and modality counts unique to the press release are not repeated on the official technical page or in the offering document.

Sources

  1. Official v1 announcement (company blog)
  2. Official v2 blog (company blog)
  3. Official v3 page (company blog)
  4. v1 press release (press)
  5. Offering document (424B4) (official site)
  6. Current report on completion of the listing (official site)

Corrections

2 corrections
FieldWasNowBasisFound
benchmarks_reportedTDC self-reported (Chinese matrix and English data)Unified to unconfirmed and not counted: the blog does not name the benchmark or give the split and protocolsource2026-11-16
latest_updatev2 (2025-05-05) as the latest event2026-09-21 v3 release; the company listing event is not counted as a model updatesource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.